Curvature and Riemann Tensor

Federico Fornari
Abstract Briey speaking this short essay has been written in order to show how, using the tools of dierential geometry, we are led to dene the machinery called Riemann tensor, an object that acts on a manifold telling us information about its curvature.

Dierentiable manifolds and tensors

The mathematical concept of a curved space begins with the idea of a manifold. A manifold is essentially a continuous space which looks locally like Euclidean space. The surface of a sphere is a manifold. So is any m-dimensional hyperplane in an n-dimensional Euclidean space (m n). Basically, a manifold is any set that can be continuously parametrized. The number of independent parameters is the dimension of the manifold, and the parameters themselves are the coordinates of the manifold. It must be stressed that the large-scale topology of a manifold may be very dierent from Euclidean space: the surface of a torus is not Euclidean, even topologically. But locally the correspondence is good: a small patch of the surface of a torus can be mapped 1-1 into the plane tangent to it. This is the way to think of a manifold: it is a space with coordinates, that locally looks Euclidean but that globally can warp, bend, and do almost anything (as long as it stays continuous). Let us consider dierentiable manifolds. These are spaces that are continuous and dierentiable. Roughly, this means it is possible to dene a scalar eld at each point of the manifold and be sure that it can be dierentiated everywhere. The surface of a sphere is dierentiable everywhere. That of a cone is dierentiable except at its apex. The assumption of dierentiability immediately means that we can dene one-forms and vectors. Using the vectors and one-forms so dened, we can build up the whole set of tensors of type m . All n of this comes only from dierentiability, so the set of all tensors is said to be part of the dierential structure of the manifold.

Riemannian manifolds

A dierentiable manifold on which a symmetric 0 tensor eld g has been sin2 gled out to act as the metric at each point is called a Riemannian manifold. It is important to understand that in picking out a metric we add structure to the manifold. Thus, by our choosing one metric g the manifold gets a certain curvature (perhaps that of a sphere), while a dierent g would give it a

dierent curvature (perhaps an ellipsoid of revolution). The dierentiable manifold itself is primitive: an amorphous collection of points, arranged locally like the points of Euclidean space, but not having any distance relation or shape specied. Giving the metric g gives it a specic shape. The metric, of course, provides a mapping between vectors and one-forms at every point. In general the components of g, g , will be complicated functions of position. Because gravity prevents inertial frames from being global, we shall have to allow all coordinates, and hence all coordinate transformations, that are nonsingular. Now, the matrix (g ) is a symmetric matrix by denition. It is a well-known theorem of matrix algebra that a transformation matrix can always be found that will make any symmetric matrix into a diagonal matrix with each entry on the main diagonal either +1, 1, or zero. The number of +1 entries equals the number of positive eigenvalues of (g ), while the number of 1 entries is the number of negative eigenvalues. So if we choose g originally to have three positive eigenvalues and one negative, then we can always nd a to make the metric components become 1 0 0 0 0 1 0 0 (1) (g ) = 0 0 1 0 = ( ). 0 0 0 1 There are two remarks that must be made here. The rst is that Eq.(1) relied on choosing g to have the appropriately signed eigenvalues. Thus, the fact that one can always construct a local inertial frame at any event, nds its mathematical representation in Eq.(1), that the metric can be transformed into at that point. The second remark is that the matrix that produces Eq.(1) at every point may not be a coordinate transformation. For instance the set of one-forms { = dx } would be a coordinate basis only if for the transformation holds: = . (2) x x In a general gravitational eld this will be impossible, because otherwise it would imply the existence of coordinates for which Eq.(1) is true everywhere: a global Lorentz frame. However, having found a basis at a particular point P for which Eq.(1) is true, it is possible to nd coordinates such that, in the neighborhood of P, Eq.(1) is nearly true. A theorem assures that if we choose any point P of the manifold a coordinate system {x } can be found whose origin is at P and in which: g (x ) = + O[(x )2 ]. (3) That is, the metric near P is approximately that of Special Relativity (SR), dierences being of second order in the coordinates. The existence of local Lorentz frames is merely the statement that any curved space has a at space tangent to it at any point.

Covariant dierentiation

We now look at the subject of dierentiation. By denition, the derivative of a vector eld involves the dierence between vectors at two dierent points (in 2

the limit as the points come together). In a curved space the notion of the dierence between vectors at dierent points must be handled with care, since in between the points the space is curved and the idea of vectors at the two points, pointing in the same direction, is fuzzy. However, the local atness of the Riemannian manifold helps us out. We only need to compare vectors in the limit as they get innitesimally close together, and we know that we can construct a coordinate system at any point which is as close to being at as we would like in this same limit. So in a small region the manifold looks at, and it is then natural to say that the derivative of a vector whose components are constant in this coordinate system is zero at that point. That is, we say that the derivatives of the basis vectors of the locally inertial coordinate system are zero at P. Let us emphasize that this is a denition of the covariant derivative. This denition immediately leads to the fact that in these coordinates at this point, the covariant derivative of a vector has components given by the partial derivatives of the components (that is, the Christoel symbols vanish): V = V at P in this frame. ; , This is of course also true for any other tensor, including the metric: g; = g, = 0 at P. (5) (4)

Now, Eq.(5) is true in one frame (the locally inertial one), and is a valid tensor equation; therefore it is true in any basis. This is a very important result, and comes directly from our denition of the covariant derivative. If we have = (it is easy to prove this for any coordinate system), then Eq.(5) leads (for any metric) to: = 1 g (g, + g, g, ). 2 (6)

We assumed at the start that at P in a locally inertial frame, = 0. But, importantly, the derivatives of at P in this frame are not all zero generally, since they involve g, . This means that even though coordinates can be found in which = 0 at a point, these symbols do not generally vanish elsewhere. This diers from at space, where a coordinate system exists in which = 0 everywhere. So we can see that at any given point, the dierence between a general manifold and a at one manifests itself in the derivatives of the Christoel symbols. Eq.(6) means that, given g , one can calculate everywhere. One can therefore calculate all covariant derivatives, given g. To review the formulas: V = V + V ; , P; = T ; = P, P T , + T T (7) (8) (9)

Parallel-transport, geodesics and curvature

It is important to distinguish two dierent kinds of curvature: intrinsic and extrinsic. Consider, for example, a cylinder. Since a cylinder is round in one 3

Figure 1: A spherical triangle ABP. direction, one thinks of it as curved. This is its extrinsic curvature: the curvature it has in relation to the at three-dimensional space it is part of. On the other hand, a cylinder can be made by rolling a at piece of paper without tearing or crumpling it, so the intrinsic geometry is that of the original paper: it is at. This means that the distance in the surface of the cylinder between any two points is the same as it was in the original paper; parallel lines remain parallel when continued; in fact, all of Euclids axioms hold for the surface of a cylinder. The intrinsic geometry of an n-dimensional manifold considers only the relationships between its points on paths that remain in the manifold. The extrinsic curvature of the cylinder comes from considering it as a surface in a space of higher dimension, and asking about the curvature of lines that stay in the surface compared with straight lines that go o it. So extrinsic curvature relies on the notion of a higher-dimensional space. When we talk about the curvature of spacetime, we talk about its intrinsic curvature, since it is clear that all world lines are conned to remain in spacetime. The cylinder, as we have just seen, is intrinsically at; a sphere, on the other hand, has an intrinsically curved surface. To see this, consider Figure 1, in which two neighboring lines begin at A and B perpendicular to the equator, and hence are parallel. When continued as locally straight lines they follow the arc of great circles, and the two lines meet at the pole P. Parallel lines, when continued, do not remain parallel, so the space is not at. There is an even more striking illustration of the curvature of the sphere. Consider, rst, at space. In Figure 2 starting at A, at each point a vector is drawn parallel to the one at the previous point. This construction is carried around the loop from A to B to C and back to A. The vector nally drawn at A is, of course, parallel to the original one. A completely dierent thing happens on a sphere! Consider the path shown in Figure 3. The vector must always be tangent to the sphere so each vector is drawn as parallel as possible to the previous one. In this loop, A and B are on the equator 90 apart, and C is at the pole. Each arc is the arc of a great circle,

Figure 2: A triangle made of curved lines in at space.

Figure 3: Parallel transport around a spherical triangle.

and each is 90 long. At A we choose the vector perpendicular to the equator. Each new vector is therefore drawn perpendicular to the arc AB. When we get to B, the vectors are tangent to BC. So, going from B to C, we keep drawing tangents to BC. These are perpendicular to the great circle of C and A, and so, from C to A the new vectors remain perpendicular to the circle. Thus the vector eld has rotated 90 in this construction! Despite the fact that each vector is drawn parallel to its neighbor, the closed loop has caused a discrepancy. Since this doesnt happen in at space, it must be an eect of the spheres curvature. This result has radical implications: on a curved manifold it simply isnt possible to dene globally parallel vector elds. One can still dene local parallelism, for instance how to move a vector from one point to another, keeping it parallel and of the same length. But the result of such parallel transport from point A to point B depends on the path taken. One therefore cannot assert that a vector at A is or is not parallel to (or the same as) a certain vector at B. The construction we have just made on the sphere is called parallel-transport. If a vector eld V is dened at every point of a curve and if the vectors V at innitesimally close points of the curve are parallel and of equal length, then d V is said to be parallel-transported along the curve. If U = d x is the tangent to the curve ( being the parameter along it; U is not necessarily normalized), then in a locally inertial coordinate system at a point P the components of V must be constant along the curve at P: dV = U V = U V = 0 at P. , ; d (10)

The rst equality is the denition of the derivative of a function (in this case V ) along the curve; the second equality comes from the fact that = 0 at P in these coordinates. But the third equality is a frame-invariant expression and holds in any basis, so it can be taken as a frame-invariant denition of the parallel-transport of V along U : UV = 0 ; dV = d
UV

= 0.

(11)

The most important curves in at space are straight lines. One of Euclids axioms is that two straight lines that are initially parallel remain parallel when extended. What does he mean by extended ? He doesnt mean continued in such a way that the distance between them remains constant, because even then they could both bend. What he means is that each line keeps going in the direction it has been going in. More precisely, the tangent to the curve at one point is parallel to the tangent at the previous point. In fact, a straight line in Euclidean space is the only curve that parallel-transports its own tangent vector! In a curved space, we can also draw lines that are as nearly straight as possible by demanding parallel-transport of the tangent vector. These are called geodesics: {U is tangent to a geodesic}
UU

=0

(12)

(Note that in a locally inertial system these lines are straight.) In component notation: U U ; = U U , + U U . (13) 6

If we let be the parameter of the curve, then U =

d d x

and U x =

d d :

dx dx d2 x + = 0. (14) d2 d d Since the Christoel symbols are known functions of the coordinates {x }, this is a nonlinear (quasi-linear), second-order dierential equation for x (). It has a unique solution when initial conditions at = 0 are given: x = x (0 ) 0 d and U0 = d x 0 . So, by giving an initial position (x ) and an initial 0 direction (U0 ), one gets a unique geodesic. Now, if is a parameter of a geodesic (so that Eq.(14) is satised), and if we dene a new parameter = a + b (15)

where a and b are constants (not depending on position on the curve), then is also a parameter in which Eq.(14) is satised. Generally speaking, only linear transformations of like Eq.(15) will give new parameters in which the geodesic equation is satised. A parameter like and above is called an ane parameter. A curve having the same path as a geodesic but parametrized by a nonane parameter is, strictly speaking, not a geodesic curve. A geodesic is also a curve of extremal length: between any two points, its length is unchanged to rst order in small changes in the curve.

The curvature tensor

At last we are in a position to give a mathematical description of the intrinsic curvature of a manifold. We go back to the curious example of the paralleltransport of a vector around a closed loop, and take it as our denition of curvature. Let us imagine in our manifold a very small closed loop (Figure 4) whose four sides are the coordinate lines x1 = a, x1 = a + a, x2 = b, and x2 = b + b. A vector V dened at A is parallel-transported to B. From the parallel-transport law e1 V = 0 we conclude V = 1 V . x1 So at B the vector has components
B

(16)

V (B) = V (A) +
A

V 1 dx = V (A) x1

x2 =b

1 V dx1

(17)

where the notation x2 = b under the integral denotes the path AB. Similar transport from B to C to D gives V (C) = V (B)
x1 =a+a

2 V dx2 1 V dx1 .

(18) (19)

V (D) = V (C) +
x2 =b+b

The integral in the last equation has a dierent sign because the direction of transport from C to D is in the negative x1 direction. Similarly, the completion of the loop gives V (Anal ) = V (D) +
x1 =a

2 V dx2 .

(20)

Figure 4: Small section of a coordinate grid. The net change in V (A) is a vector V , found by adding Eqs.(17)-(20): V = V (Anal ) V (Ainitial ) = =
x1 =a

2 V dx2

x1 =a+a

2 V dx2 + (21)

+
x2 =b+b

1 V dx1

x2 =b

1 V dx1 .

Notice that these would cancel in pairs if and V were constants on the loop, as they would be in at space. But in curved space they are not, and we get to lowest order
b+b

b a+a

a b

2 V dx2 + x1

1 V dx1 x2 a 1 V a b 1 2 V + x x2

(22)

This involves derivatives of Christoel symbols and of V . The derivatives V can be eliminated using Eq.(16) and its equivalent with 1 replaced by 2. Then Eq.(22) becomes V = a b 1,2 2,1 + 2 1 1 2 V 8 (23)

(To obtain this, one needs to relabel dummy indices in the terms quadratic in s.) Now, the indices 1 and 2 appear because the path was chosen to go along those coordinates. It is antisymmetric in 1 and 2 because the change V would have to have the opposite sign if one went around the loop in the opposite direction (that is, interchanging the roles of 1 and 2). If one used general coordinate lines x and x , one would nd V = change in V due to transport, rst a e , then b e , then a e , and nally b e = = a b , , + V . (24)

Now, V depends on a b, the coordinate area of the loop. So it is clear that if the length of the loop in one direction is doubled, V is doubled. This means that V depends linearly on a e and b e . Moreover, it certainly also depends linearly in Eq.(24) on V itself and on , which is the basis one-form that gives V from the vector V . Hence we have the following result: if we dene R = , , + (25) then R must be components of the 1 tensor which, when supplied with 3 arguments , V , a e , b e , gives V , the component of the change in V on parallel-transport around a loop given by a e and b e . This tensor is called the Riemann curvature tensor R. It is useful to look at the components of R in a locally inertial frame at a point P. We have = 0 at P but from Eq.(6) , = 1 g (g, + g, g, ). 2 (26)

Since second derivatives of g dont vanish, we get at P R = 1 g (g, + g, g, g, g, + g, ). 2 (27)

Using the symmetry of g and the fact that g, = g, because partial derivatives always commute, we nd R = 1 g (g, g, + g, g, ). 2 (29) (28)

If we lower the index we get (in this coordinate system at P) R = g R = 1 (g, g, + g, g, ). 2 (30)

In this form it is easy to verify the following identities: R = R = R = R R + R + R = 0. (31) (32)

Thus, R is antisymmetric on the rst pair and on the second pair of indices, and symmetric on exchange of the two pairs. Since Eqs.(31) and (32) are valid 9

tensor equations true in one coordinate system, they are true in all bases. (Note that an equation like Eq.(29) is not a valid tensor equation, since it involves partial derivatives, not covariant ones. Therefore it is true only in the coordinate system in which it was derived.) It can be shown that the various identities, Eqs.(31) and (32), reduce the number of independent components of R (and hence of R ) to 20, in four dimensions. This is, not coincidentally, the same number of independent g, that could not be made to vanish by a coordinate transformation (Eq.(3)). Thus R characterizes the curvature in a tensorial way. A at manifold is one which has a global denition of parallelism: a vector can be moved around parallel to itself on an arbitrary curve and will return to its starting point unchanged. This clearly means that R = 0 at manifold. (33)

An important use of the curvature tensor comes when we examine the consequences of taking two covariant derivatives of a vector eld V . We found that rst derivatives were like at-space ones, since we could nd coordinates in which the metric was at to rst order. But second derivatives are a dierent story:
V

V = V ; ;

+ V ; V . ;

(34)

In locally inertial coordinates at P, all the s are zero, but their partial derivatives are not. Therefore we have at P
V

= V + , V ,

(35)

in these coordinates only. Consider the same formula with and exchanged:
V

= V + , V . ,

(36)

If we subtract these we get the commutator of the covariant derivative operators and , written in the same notation as we would employ in quantum mechanics: [
, ] V

= , , V .

(37)

The terms involving the second derivatives of V drop out here, since V =V , , (38)

(Let us pause to recall that V is the partial derivative of the component V , , so by the laws of partial dierentiation the partial derivatives must commute. On the other hand, V is a component of the tensor V , and V is a component of V : there is no reason from dierential calculus why it must be symmetric on and .) Now, in this frame (where = 0), we can compare Eq.(37) with Eq.(25) and see that at P [
, ] V

= R V .

(39)

Since this is a valid tensor equation, it is true in any coordinate system. The Riemann tensor gives the commutator of covariant derivatives. This means 10

that in curved spaces, one must be careful to know the order in which covariant derivatives are taken: they do not commute. This can be extended to tensors of higher rank. For example, a 1 tensor has 1 [
, ] F

= R F + R F .

(40)

That is, each index gets a Riemann tensor on it, and each one comes in with a + sign. (They must all have the same sign because raising and lowering indices with g is unaected by , since g = 0.) Eq.(39) is closely related to our original derivation of the Riemann tensor from parallel-transport around loops, because the parallel-transport problem can be thought of as computing, rst the change of V in one direction, and then in another, followed by subtracting changes in the reverse order: this is what commuting covariant derivatives also does. Indeed we made our derivation of the Riemann tensor by calculating the change in a vector parallely transported along a closed loop generated by a set of coordinate basis vectors; but what happens if we parallely transport the same vector along the closed path formed by two generical vector elds? Let us d d consider at a point P the vector elds V = d and W = d , and a vector A (in the tangent space we dene the coordinate vectors e = x ). Recalling Eq.(13), the covariant derivative of V along W takes the form
WV

= W V + V W e = ,

d V + V W e d

(41)

and similarly (interchanging the roles of V and W )

V

W =

d W + W V e . d

(42)

Again we subtract one equation from the other. After relabeling dummy indices this yields
WV

W =

d d V W e + V W e . d d

(43)

In any coordinate system we know that = . Thus if we dene the vector elds commutator (which as we can see is still a vector) W,V = the former expression becomes
WV

d d V W e d d

(44)

W = W,V .

(45)

From a pictorial point of view it is immediate to appreciate the meaning of this relation (Figure 5). The commutator represents the vector which closes the path left open by the transport of each vector eld along the other ones integral curve. Therefore, if we want to parallely transport A along a loop in order to quantify its net change, we must repeat the construction previously shown within Eqs.(16)-(24) taking in account that this time A = change in A due to transport, rst V , then W , then W , V , then V , and nally W . (46)

11

Figure 5: Symmetric covariant dierentiation and pictorial view of the commutator. For this reason, remembering what we obtained in Eq.(24) and Eq.(39), we are led to re-dene the Riemann tensor in a more general way as follows R V W A e =
V

[V ,W ] A. = W,V

(47) = 0

Note that if V and W are coordinate vectors then V , W

(actually bringing us back to Eq.(39)). We have often mentioned that in a curved space, parallel lines when extended do not remain parallel. This can now be formulated mathematically in terms of the Riemann tensor. Consider two geodesics (with tangents V and V ) that begin parallel and near each other, as in Figure 6, at points A and A . Let the ane parameter on the geodesics be called . We dene a connecting vector which reaches from one geodesic to another, connecting points at equal intervals in (i.e. A to A , B to B , etc.). For simplicity, let us adopt a locally inertial coordinate system at A, in which the coordinate x0 points along the geodesics. Thus at A we have V = 0 . The equation of the geodesic at A is d2 x =0 (48) d2 A since all Christoel symbols vanish at A. The Christoel symbols do not vanish at A , so the equation of the geodesic V at A is d2 x d2 + 00 (A ) = 0
A

(49)

where again at A we have arranged the coordinates so that V = 0 . But,

12

Figure 6: A connecting vector between two geodesics connects points of the same parameter value. since A and A are separated by , we have 00 (A ) 00, = the right-hand side being evaluated at A. With Eq.(49) this gives d2 x d2 = 00, .
A

(50)

(51)

Now, the dierence x (, geodesic V )x (, geodesic V ) is just the component of the vector . Therefore, at A, we have d2 = d2 d2 x d2
A

d2 x d2

= 00, .
A

(52)

This then gives how the components of change. But since the coordinates are to some extent arbitrary, we want to have, not merely the second derivative of the component , but the full second covariant derivative V V . We can use Eq.(11) to obtain
V V

) =

d ( d

) + 0

(53)

Now, using = 0 at A, we have

V V

d d

d + 0 d 13

+0=

d2 + 0,0 d2

(54)

at A. (We have also used ,0 = 0 at A, which is the condition that curves begin parallel.) So we get
V V

= 0,0 00, = R00 = R V V

(55)

where the second equality follows from Eq.(25). The nal expression is frame invariant, so we have, in any basis,
V V

= R V V .

(56)

Geodesics in at space maintain their separation; those in curved spaces dont. This is called the equation of geodesic deviation and shows mathematically that the tidal forces of a graviational eld (which cause trajectories of neighboring particles to diverge) can be represented by curvature of a spacetime in which particles follow geodesics.

Bianchi identities, Ricci and Einstein tensors

Let us return to Eq.(25) for the Riemann tensors components. If we dierentiate it with respect to x (just the partial derivative) and evaluate the result in locally inertial coordinates, we nd R, = 1 (g, g, + g, g, ). 2 (57)

From this equation, the symmetry g = g and the fact that partial derivatives commute, one can show that R, + R, + R, = 0. (58)

Since in our coordinates = 0 at this point, this equation is equivalent to R; + R; + R; = 0. (59)

But this is a tensor equation, valid in any system. It is called the Bianchi identities and is very important for General Relativity (GR). Before pursuing the consequences of the Bianchi identities, we shall need to dene the Ricci tensor R : R R = R . (60) It is the contraction of R on the rst and third indices. Other contractions would in principle also be possible: on the rst and second, the rst and fourth, etc. But because R is antisymmetric on and and on and , all these contractions either vanish identically or reduce to R . Therefore the Ricci tensor is essentially the only contraction of the Riemann tensor. Note that Eq.(31) implies it is a symmetric tensor. Similarly, the Ricci scalar is dened as R g R = g g R . (61) Let us apply the Ricci contraction to the Bianchi identities, Eq.(59): g [R; + R; + R; ] = 0

14

or R; R; + R; = 0. (62) To derive this result one needs two facts. First, by Eq.(5) we have g; = 0; since g is a function only of g it follows that g ; = 0. Therefore g and g can be taken in and out of covariant derivatives at will: index-raising and -lowering commutes with covariant dierentiation. The second fact is that g R; = g R; = R; (63)

accounting for the second term in Eq.(62). Eq.(62) is called the contracted Bianchi identities. A more useful equation is obtained by contracting again on the indices and : g R; R; + R; = 0 or R; R; R; = 0. (64) Again the antisymmetry of R has been used to get the correct sign in the last term. Note that since R is a scalar, R; R, in all coordinates. Now, Eq.(64) can be written in the form 2R R = 0.
;

(65)

These are the twice-contracted Bianchi identities, often simply also called the Bianchi identities. If we dene the symmetric tensor 1 G R g R = G 2 then we see that Eq.(65) is equivalent to G ; = 0. (67) (66)

The tensor G is constructed only from the Riemann tensor and the metric, and is automatically divergence free as an identity. It is called the Einstein tensor, since its importance for gravity was rst understood by Einstein. In fact the Einstein eld equations for GR are G = 8GN T (68)

(where T is the stress-energy tensor and GN is the universal gravitational constant rst introduced by Newton). The Bianchi identities then imply T ; = 0 which is the equation of local conservation of energy and momentum. (69)

References
[1] B. F. Schutz: A rst course in general relativity, Cambridge University Press, Chapter 6 (85). [2] C. W. Misner, K. S. Thorne, J. A. Wheeler: Gravitation, W. H. Freeman and Company, S. Francisco, Chapter 10 (73).

15